Windows Alpha 2 for AMD & Nvidia GTX GPUs only

suraj · February 22, 2023, 8:49pm

Hi All,

Here is a version of the app with all onnx specific optimizations complete. This version should be almost 2X faster and should work for all Nvidia(non RTX) and AMD GPUs. Please let us know if there is no speed up or drop in performance.

Download: Win

Known Issues:

Proteus Auto will not work, use Proteus with manual settings
Models cannot be run on CPU
Models cannot be run on intel GPU
Models run slower on RTX GPUs

Thank you for testing, let us know if you have any issues.

PS: Since this is an alpha version, it will not interfere with Beta or Release versions on the machine.

Martyprod · February 22, 2023, 9:18pm

1/ no speed improvement here. same speed as before on 1050Ti GTX. 0.78s/f for a 720p->1080p. Ram usage in preference at 80% or 100% give the same result. i’ll test an another file not on my Nas.
2/ bug on pictures in the preview. the bug is there on the external video player too. didn’t do an export yet to see if there is a difference. see below :

Martyprod · February 22, 2023, 9:21pm

another file, local drive, (the other one was local too), but in 640x480 → 1080p. the alpha doesn’t like Harrisson ford lol no improvement on speed.

suraj · February 22, 2023, 9:25pm

Can you send me a few log files
What is the GPU utilization for the current release version, is it close to 100%? If yes then there will be no speed up.

Martyprod · February 22, 2023, 9:45pm

yes, i can send you the logs i tried i think with 2 different % ram usage (100% and 80%) but i’ll retry again. in the MSI afterburner, Gpu was at 100%)

Martyprod · February 22, 2023, 9:51pm

Max memory usage : 80%
Windows report Gpu usage : 100%
Speed improvement : No (same 0.78spf)
bug on picture. (logs is coming).

david.123 · February 22, 2023, 9:52pm

Not seeing any difference in performance on GTX card either. Is there a specific model or scenario where we can expect to see an improvement? Or is this mainly for AMD GPUs?

Martyprod · February 22, 2023, 9:53pm

3 Logs sent in PM through the forum.

tried at 70%, no speed improvement, graphic bug still there. last Nvidia driver studio version.

Tried the low power mode in preference, the Gpu in windows now show 95% (but anyway, i limit the use of my graphic card in MSI afterburner at 92% so it’s never 100%). speed decresed from 0.78 to 0.77 !! we’re close !! it’s speed up !
95% in Preference of Ram usage + low power mode = decrease to 0.76.

TPX · February 22, 2023, 10:49pm

Just a short test on my Intel RTX sys, my processor almost hit its overheating point, but had the fastest time with 27 sec (25 for second run) for a pal to hd conversion with 27 sec. (10 sec sequence) (before was 30 on this sys).

I’ve black bars also when i use a memory reduction.
There also filckering artifacts.

Screenshot 2023-02-22 235502

2023-02-22-23-41-29-Main.txt (657.5 KB)

Will test tomorrow on my AMD sys.

Martyprod · February 22, 2023, 10:56pm

which graphic card did you used or Cpu ?

TPX · February 22, 2023, 11:02pm

7820X & RTX Quadro 5000.

Martyprod · February 22, 2023, 11:09pm

ok, so not concerned by this alpha, as it seems it’s for GTX and AMD card. but despite this, it speed up your processing by using it with an RTX compared to the 3.1.6 normal release ?
i’ll try, i have a cloud computer account with a A4500 Quadro RTX. close to you (almost).

suraj · February 22, 2023, 11:21pm

The version is for all windows GPUs. RTX cards are running TensorRT models so I don’t expect this version to be faster.

@Martyprod Do you still see artifacts when running in low power mode?

The 2X speed up numbers are from a AMD W6800 and 7900XTX for 1080p to 4K using Artemis models. On the RTX 2070 & 3090 the performance is similar to release version.

Unfortunately, I don’t have access to many older GPUs, so it will be hard to fix artifact issues. Expect more versions.

Martyprod · February 22, 2023, 11:30pm

Yes, there is still the artificats (big ones).

No problem to test more versions, if it can be fixed, it would be more than welcome if it can increase the speed for Gtx card too.

TPX · February 23, 2023, 7:48am

Could you erase your models in the alpha model folder, and retest?

TPX · February 23, 2023, 7:56am

Preliminary:
The performance increase is there, up to 2x, but here and there there are problems.

I got 30 fps (only in low power mode ² ) with my pal to FHD test video, before was 10 - 12 fps, so almost 3x speedup, but the output is broken.

But it seems like its only broken with the small models.

“prob-v3-fgnet-fp16-576x384-2x-ox” ← has artifacts, colored blotches
“prob-v3-fgnet-fp16-448x576-4x-ox” ← is broken, black boxes
“alq-v13-fgnet-fp16-448x576-4x-ox” ← has tile artifact
“amq-v13-fgnet-fp16-448x576-4x-ox” ← has tile artifact
“prob-v3-fgnet-fp16-448x576-1x-ox” ← is broken, black boxes
“thf-v4-fnet-fp16-448x576-ox” ← flickering artifacts + no speedup to 3.1.6 and has artifacts

We need a display which models are currently used, this would speed up everything, this is so time-consuming, this would also simplify debugging in the forum.

I did not test all models and not all are broken.

The larger the models in the resolution, the less likely it is that they are defective.

When i try to upscale a pal video to 8K i get this error:

Last FFmpeg messages:
encoder->Init() failed with error 5
Error initializing output stream 1:0 – Error while opening encoder for output stream #1:0 - maybe incorrect parameters such as bit_rate, rate, width or height
Qavg: 16073.476
2 frames left in the queue on closing

Low Power Mode: ²
After switching back to normal the performance did increase.

After further tests I found that sometimes it makes a difference, but sometimes not.

Alpha 3.1.7.1

Proteus & Artemis low

Stable 3.1.6

Proteus & Artemis low

Martyprod · February 23, 2023, 9:39am

Yes I can, will test immediatly.

Martyprod · February 23, 2023, 10:02am

it loaded the prob-v3-fgnet-fp32-384x672-2x-ox.tz model, and same issue.
artifacts, two big black bars on the preview. speed is 0.75 spf per frame for 720p/60 to 1080p/60 using low power mode + 95% ram usage.
speed was 0.76 exterday. normal 3.1.6 release give a 0.78spf.
(didn’t tested yet on the A4500RTX / AMD Epyc system).
Without power mode, it goes back to 0.78.

trying with an another video 720p/50 to 1080p/50. the estimate button is not working by the way.
same issue, speed, and model used

i’ll try some SD square stuff.

TPX · February 23, 2023, 10:09am

You just need to change the upscale to get the bigger models, seems like the small ones are broken.